1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * (C) 2008-2009 Pablo Neira Ayuso <pablo@netfilter.org>
4 */
5#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
6#include <linux/module.h>
7#include <linux/skbuff.h>
8#include <linux/jhash.h>
9#include <linux/ip.h>
10#include <net/ipv6.h>
11
12#include <linux/netfilter/x_tables.h>
13#include <net/netfilter/nf_conntrack.h>
14#include <linux/netfilter/xt_cluster.h>
15
16static inline u32 nf_ct_orig_ipv4_src(const struct nf_conn *ct)
17{
18	return (__force u32)ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip;
19}
20
21static inline const u32 *nf_ct_orig_ipv6_src(const struct nf_conn *ct)
22{
23	return (__force u32 *)ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.ip6;
24}
25
26static inline u_int32_t
27xt_cluster_hash_ipv4(u_int32_t ip, const struct xt_cluster_match_info *info)
28{
29	return jhash_1word(ip, info->hash_seed);
30}
31
32static inline u_int32_t
33xt_cluster_hash_ipv6(const void *ip, const struct xt_cluster_match_info *info)
34{
35	return jhash2(ip, NF_CT_TUPLE_L3SIZE / sizeof(__u32), info->hash_seed);
36}
37
38static inline u_int32_t
39xt_cluster_hash(const struct nf_conn *ct,
40		const struct xt_cluster_match_info *info)
41{
42	u_int32_t hash = 0;
43
44	switch(nf_ct_l3num(ct)) {
45	case AF_INET:
46		hash = xt_cluster_hash_ipv4(nf_ct_orig_ipv4_src(ct), info);
47		break;
48	case AF_INET6:
49		hash = xt_cluster_hash_ipv6(nf_ct_orig_ipv6_src(ct), info);
50		break;
51	default:
52		WARN_ON(1);
53		break;
54	}
55
56	return reciprocal_scale(hash, info->total_nodes);
57}
58
59static inline bool
60xt_cluster_is_multicast_addr(const struct sk_buff *skb, u_int8_t family)
61{
62	bool is_multicast = false;
63
64	switch(family) {
65	case NFPROTO_IPV4:
66		is_multicast = ipv4_is_multicast(ip_hdr(skb)->daddr);
67		break;
68	case NFPROTO_IPV6:
69		is_multicast = ipv6_addr_is_multicast(&ipv6_hdr(skb)->daddr);
70		break;
71	default:
72		WARN_ON(1);
73		break;
74	}
75	return is_multicast;
76}
77
78static bool
79xt_cluster_mt(const struct sk_buff *skb, struct xt_action_param *par)
80{
81	struct sk_buff *pskb = (struct sk_buff *)skb;
82	const struct xt_cluster_match_info *info = par->matchinfo;
83	const struct nf_conn *ct;
84	enum ip_conntrack_info ctinfo;
85	unsigned long hash;
86
87	/* This match assumes that all nodes see the same packets. This can be
88	 * achieved if the switch that connects the cluster nodes support some
89	 * sort of 'port mirroring'. However, if your switch does not support
90	 * this, your cluster nodes can reply ARP request using a multicast MAC
91	 * address. Thus, your switch will flood the same packets to the
92	 * cluster nodes with the same multicast MAC address. Using a multicast
93	 * link address is a RFC 1812 (section 3.3.2) violation, but this works
94	 * fine in practise.
95	 *
96	 * Unfortunately, if you use the multicast MAC address, the link layer
97	 * sets skbuff's pkt_type to PACKET_MULTICAST, which is not accepted
98	 * by TCP and others for packets coming to this node. For that reason,
99	 * this match mangles skbuff's pkt_type if it detects a packet
100	 * addressed to a unicast address but using PACKET_MULTICAST. Yes, I
101	 * know, matches should not alter packets, but we are doing this here
102	 * because we would need to add a PKTTYPE target for this sole purpose.
103	 */
104	if (!xt_cluster_is_multicast_addr(skb, xt_family(par)) &&
105	    skb->pkt_type == PACKET_MULTICAST) {
106	    	pskb->pkt_type = PACKET_HOST;
107	}
108
109	ct = nf_ct_get(skb, &ctinfo);
110	if (ct == NULL)
111		return false;
112
113	if (ct->master)
114		hash = xt_cluster_hash(ct->master, info);
115	else
116		hash = xt_cluster_hash(ct, info);
117
118	return !!((1 << hash) & info->node_mask) ^
119	       !!(info->flags & XT_CLUSTER_F_INV);
120}
121
122static int xt_cluster_mt_checkentry(const struct xt_mtchk_param *par)
123{
124	struct xt_cluster_match_info *info = par->matchinfo;
125	int ret;
126
127	if (info->total_nodes > XT_CLUSTER_NODES_MAX) {
128		pr_info_ratelimited("you have exceeded the maximum number of cluster nodes (%u > %u)\n",
129				    info->total_nodes, XT_CLUSTER_NODES_MAX);
130		return -EINVAL;
131	}
132	if (info->node_mask >= (1ULL << info->total_nodes)) {
133		pr_info_ratelimited("node mask cannot exceed total number of nodes\n");
134		return -EDOM;
135	}
136
137	ret = nf_ct_netns_get(par->net, par->family);
138	if (ret < 0)
139		pr_info_ratelimited("cannot load conntrack support for proto=%u\n",
140				    par->family);
141	return ret;
142}
143
144static void xt_cluster_mt_destroy(const struct xt_mtdtor_param *par)
145{
146	nf_ct_netns_put(par->net, par->family);
147}
148
149static struct xt_match xt_cluster_match __read_mostly = {
150	.name		= "cluster",
151	.family		= NFPROTO_UNSPEC,
152	.match		= xt_cluster_mt,
153	.checkentry	= xt_cluster_mt_checkentry,
154	.matchsize	= sizeof(struct xt_cluster_match_info),
155	.destroy	= xt_cluster_mt_destroy,
156	.me		= THIS_MODULE,
157};
158
159static int __init xt_cluster_mt_init(void)
160{
161	return xt_register_match(&xt_cluster_match);
162}
163
164static void __exit xt_cluster_mt_fini(void)
165{
166	xt_unregister_match(&xt_cluster_match);
167}
168
169MODULE_AUTHOR("Pablo Neira Ayuso <pablo@netfilter.org>");
170MODULE_LICENSE("GPL");
171MODULE_DESCRIPTION("Xtables: hash-based cluster match");
172MODULE_ALIAS("ipt_cluster");
173MODULE_ALIAS("ip6t_cluster");
174module_init(xt_cluster_mt_init);
175module_exit(xt_cluster_mt_fini);
176